Early screening for autism spectrum disorder (ASD) is crucial, yet current assessment tools in Chinese primary child care are limited in efficacy. This study aims to employ machine learning algorithms to identify key indicators from the 20-item Modified Checklist for Autism in Toddlers, revised (M-CHAT-R) combining with ASD-related sociodemographic and environmental factors, to distinguish ASD from typically developing children. Data from our prior validation study of the Chinese M-CHAT-R (August 2016-March 2017, n = 6,049 toddlers) were reviewed. We extracted the 20-item M-CHAT-R data and integrated 17 sociodemographic and environmental risk factors associated with ASD development to strengthen M-CHAT-R's machine learning screening. Five feature selection methods were used to extract subsets from the original set. Six machine learning algorithms were applied to identify the optimal subset distinguishing clinically diagnosed ASD toddlers from typically developing toddlers. Nine features were grouped into three subsets: subset 1 contained unanimously recommended items (A1 [Follows point], A3 [Pretend play], A9 [Brings objects to show], A10 [Response to name] and A16 [Gazing following]). Subset 2 added two items (A17 [Gaining parent's attention] and A18 [Understands what is said]), and subset 3 included two more items (A8 [Interest in other children] and child's age). The top-performing algorithm resulted in a seven-item classifier of subset 2 with 92.5 % sensitivity, 90.1 % specificity, and 10.0 % positive predictive value. Machine learning classifiers effectively differentiate ASD toddlers from typically developing toddlers using a reduced M-CHAT-R item set. This highlights the clinical significance of machine learning-optimized models for ASD screening in primary health care centers and broader applications.
Read full abstract